Shoko
Araki
NTT Communication Science Laboratories
shoko@cslab.kecl.ntt.co.jp
*** ALGORITHM ***
(Source number estimation + TDOA clustering) + time-frequency masking. 

Source number estimation and clustering are done simultaneously by TDOA GMM fitting where a prior model is given to the mixing weight. GMM parameters are estimated by the EM algorithm. Initially 8 Gaussians are prepared. After the fitting, some Gaussians are dominant. We consider each resultant dominant Gaussian corresponds to each source. 
The time-frequency masks are given by corresponding Gaussian.

The basic idea using a prior information is submitted to the ICASSP2009 (S. Araki, T. Nakatani, H. Sawada and S. Makino, \"Blind sparse source separation for unknown number of sources using Gaussian mixture model fitting with Dirichlet prior\"), and its extention to a spatial aliasing case will be submitted to ICA2009, hopefully. 

*** GIVEN Info ***
The microphone spacing is given (but not mandatory). The number of sources is NOT given.

*** COMPUTATIONAL TIME ***
I just tried \"_liverec_*_5cm\" case. 
With Athron64 24GHz, the time was around 50 seconds.  (10 EM iterations)